bak

10-Qs for fun. Big Tech, AI infra, IPOs, macro. Notes.

[bak @copiumfueled]

David Sacks asked the question everyone building data centers keeps dodging: shouldn't this stuff be getting cheaper over time?

Gavin Baker answered with a number.

The assumption under the whole buildout is that compute gets cheaper. Chips improve, scale kicks in, cost per unit falls.

Baker says the opposite is happening right now.

A single gigawatt data center is 35 billion dollars in Nvidia silicon and 25 billion dollars in power and cooling. That second number is the problem.

A big chunk of it is human labor to install, and labor doesn't follow Moore's law. It goes up.

Then there's memory. DRAM is heading toward 30 to 40 percent of all hyperscaler capex next year.

Hundreds of billions flowing into a component that only three companies on earth know how to make, and bringing new capacity online takes years.

So the cost to stand up a gigawatt is climbing, not falling.

This is what gets lost when people watch revenue catch up to depreciation.

Depreciation is calculated on what you already spent, at prices you already paid. It's a number from the past.

The cost of the next gigawatt is a number from the future, and that one keeps moving.

Revenue can grow fast enough to close that gap. It won't matter much if the gap keeps widening from the other side.

Watch the full episode on @theallinpod

[Oguz Erkan @oguzerkan]

Bloomberg: "AI demand begins to justify data center buildout."

Based on the current cost of capital and operating margins of hyperscalers and the depreciation periods, ROI on AI capex turns positive at ~1.7-1.8x revenue/D&A.

We are now around 1.2x.

If capex doesn't explode and sales growth remains robust, ROI will likely turn positive somewhere in the next 24 months.

[bak @copiumfueled]

https://t.co/CNVEONp5Ja

[bak @copiumfueled]

More on this: Politico reports the Trump administration plans to lift export restrictions on Fable 5 as soon as Tuesday evening. It's a full re-release for all users, not just the US.

[bak @copiumfueled]

Looks like Fable is coming back.

[bak @copiumfueled]

Looks like Fable is coming back.

[Sophia Cai @SophiaCai99]

🚨 NEWS: Commerce is expected to lift export controls on Fable tonight, a senior White House official tells me.

[bak @copiumfueled]

https://t.co/FlurEzbsbc

[bak @copiumfueled]

Satya Nadella explained why renting the best AI model costs a company its edge at the frontier.

And why that decides who ends up owning the frontier itself.

If a model learns from data, he asked, what is the future of the firm? A firm today is its private knowledge and its people.

Feed that into someone else's model and the value walks out with the tokens.

So Microsoft licenses its models with the weights included. Any company can set up its own training environment, its own private evals, and let a model train on its own data.

The traces, the IP, the outcomes stay with the company.

He was specific about what compounds. Not just human capital. Token capital.

Then the line that matters: if you're just a consumer of a foundation model, Nadella isn't sure how you keep enterprise value, let alone create it.

And that's his actual point. The only way the whole thing stays positive-sum, with lots of companies at the frontier instead of a few labs taking everything, is if every company builds and retains its own IP.

Full talk on @StanfordOnline

[bak @copiumfueled]

"Human Capital, AI, and Labor Commoditization," Auyon Siddiq, Niuniu Zhang, UCLA Anderson, June 2026.

arxiv. org/abs/2606.21880

[bak @copiumfueled]

Mark Zuckerberg laid out his bet on how AI will hit employment.

He acknowledges that companies always automate, and if automation outpaces the growth in people's productivity, jobs can shrink.

But he doesn't see that as inevitable. He sees it as a question of balance. His conclusion: if you point the technology at making each person more productive (that's his idea of "personal superintelligence": AI that works for the individual user), then jobs grow rather than shrink.

For him it all comes down to empowering the individual.

This is how Zuckerberg sees it. But there's another angle, one that comes from data on a market already up and running.

A new study by Auyon Siddiq and Niuniu Zhang (UCLA Anderson) finds the opposite playing out in one real-world market.

They took Upwork data and compared what drives demand for a worker: their human capital (experience, skills, reputation, portfolio) or price.

The comparison was built around the release of ChatGPT, before and after its release, across categories with different exposure to AI.

What they found: in categories more exposed to AI, human capital matters less (about 7.8% less) and price matters more.

The premium for strong human capital shrinks, and demand shifts toward cheaper workers. The effect grows over time. The authors call it the commoditization of labor.

So far, in this market, AI isn't lifting the standout specialist. It's making labor more interchangeable.

Experience and reputation count for less, price for more.

That runs counter to Zuckerberg's bet.

Watch the full conversation on @Complex

[bak @copiumfueled]

If the model isn't the moat, the question one layer down gets even sharper. Someone still has to build the gigawatt, and that capex doesn't compress on a 2-3-month cycle the way the premium does. So you've got value collapsing toward free at the top while the cost to stand up the next data center keeps climbing. Curious where the margin actually lands once both of those play out

[bak @copiumfueled]

Yeah, I'm with you on the value trajectory. That part I don't doubt. The piece i keep turning over is who captures it. Even if you own the whole stack, if token cost drop orders of magnitude like you said, hyperscalers competing on price hand a lot of that value straight to the user. And Nvidia's already taken its cut up front. Sand and sunlight are abundant. HBM and skilled install labor aren't, at least not yet.

[bak @copiumfueled]

David Sacks asked the question everyone building data centers keeps dodging: shouldn't this stuff be getting cheaper over time?

Gavin Baker answered with a number.

The assumption under the whole buildout is that compute gets cheaper. Chips improve, scale kicks in, cost per unit falls.

Baker says the opposite is happening right now.

A single gigawatt data center is 35 billion dollars in Nvidia silicon and 25 billion dollars in power and cooling. That second number is the problem.

A big chunk of it is human labor to install, and labor doesn't follow Moore's law. It goes up.

Then there's memory. DRAM is heading toward 30 to 40 percent of all hyperscaler capex next year.

Hundreds of billions flowing into a component that only three companies on earth know how to make, and bringing new capacity online takes years.

So the cost to stand up a gigawatt is climbing, not falling.

This is what gets lost when people watch revenue catch up to depreciation.

Depreciation is calculated on what you already spent, at prices you already paid. It's a number from the past.

The cost of the next gigawatt is a number from the future, and that one keeps moving.

Revenue can grow fast enough to close that gap. It won't matter much if the gap keeps widening from the other side.

Watch the full episode on @theallinpod

[Oguz Erkan @oguzerkan]

Bloomberg: "AI demand begins to justify data center buildout."

Based on the current cost of capital and operating margins of hyperscalers and the depreciation periods, ROI on AI capex turns positive at ~1.7-1.8x revenue/D&A.

We are now around 1.2x.

If capex doesn't explode and sales growth remains robust, ROI will likely turn positive somewhere in the next 24 months.

[bak @copiumfueled]

China's AI playbook, or Washington's own misfire?

Aaron Levie went on CNBC this week and read it from the enterprise-software side.

For years, AI shipped like any other software export, no questions asked.

Now access has a switch, and the precedent is set.

Cheap intelligence looks like a price story. Levie read it as a story about access.

The cheaper the intelligence gets, the more control over it is worth.

When Anthropic cut off access to Fable Mythos, its most powerful model, policymakers around the world asked one question: how do you build on a model when access can be cut off at any moment?

Levie said the real misplay may have been Washington's, not Anthropic's.

The first country to block an AI model through export controls just made sovereign AI a requirement for everyone else.

If you run the EU, you don't need a theory about Beijing. You need a backstop. You keep working with US labs, but now you fund your own open weights too, because you've seen what dependence costs once it turns into leverage.

Picture it the way Levie framed it: AI as a lever in a trade negotiation five years out. A year ago he wouldn't have put "block access to a model" on the board of possible chess moves. Now it's been played. And it doesn't get un-played.

So the cheap intelligence everyone's reading as a price story traces back to a switch the West flipped itself.

The most open country in the race was the first to prove access could be cut.

Everyone else took notes.

Watch the full conversation on @CNBC

[Xiaoyin Qu @quxiaoyin]

China’s AI playbook: kill OpenAI and anthropic with free great models. Make it free. Then use cheap electricity to export compute as well. Currently the blocker is chip but Hauwei would catch up soon. Imagine a world where instead of paying hundreds of billions to OpenAI and anthropic, you pay almost zero to similar level of intelligence with cheap cheap inference. What’s gonna happen?

[bak @copiumfueled]

@Pandering_Panda Agreed. The post put Dario at the start of the chain, not the end of it. The administration had the kill switch and used it. Sacks said it himself: "we invented reasons not to ship." We. Not Dario.

[bak @copiumfueled]

The two best AI models in America are sitting in a drawer right now.

David Sacks explained why this week on All-In, from inside the administration.

China just dropped GLM 5.2. Open weight, MIT license, free to download anywhere. It beats GPT 5.5 on coding and trails Claude Opus 4.8 by under a point.

The best openweight model on earth, and it came out of Beijing.

Fable got rolled back. GPT 5.6 is stuck navigating new approval hoops. Both frozen by a jailbreak report and a safety fight nobody outside Washington fully understands yet.

So the most open country ended up with the closed models, and right now they are off the market.

Sacks traced how America got here. Dario spent months lobbying for a federal AI regulator. He wrote a blog post asking for an FAA for AI. A government approval process for model releases. And he got it.

The administration moved, and the first thing the new caution did was freeze his own company's model.

The Chinese model has no such problem. No approval queue. No regulator. Trained on Huawei chips, packaged as "AI in a box," sold globally at a fraction of the cost.

Sacks has been saying it for months. We invented reasons not to ship. We defined this as a race and then handed the other side a head start.

What everyone is pointing at is real, but none of this was fate. A freer society didn't take the open models. America shut its own best models in a room and called it safety.

The closed models come from the open nation because the open nation chose to close them.

Watch the whole thing on @theallinpod

[Boring_Business @BoringBiz_]

The irony of the open source AI models coming from a closed society, while the closed models come from the open democratic nation, is not talked about enough

[bak @copiumfueled]

@InviertoConIA Fair, that's the honest bull case.But railroads and fiber were land grabs too, the winners won huge and most overbuilders went bankrupt before demand showed up. Capex keeps climbing while you wait. Most don't survive the gap.

[bak @copiumfueled]

You're right the tax side already tilts hard toward capital. Full expensing came back under OBBBA, and the scarce inputs price like bottlenecks. I'd just separate the mechanics from the fairness. The cost base rises because the slow inputs gain pricing power whoever ends up capturing the rents. Who bears it, and whether that's fair, is a separate fight from whether the next gigawatt costs more.

[bak @copiumfueled]

That's a sharp framing, and close. It's Baumol plus a supply crunch. The compute side improves on a Moore's law curve, but pouring pads, running powerr, and plumbing cooling don't, so as the fast inputs get cheaper the slow ones eat more of the bill. Pure Baumol is just wage pull from the productive sector. Here you've also got physical scarcity, transformers on multi-year lead times, grid interconnect queues, three vendors for memory. Those slow inputs aren't holding steady. They're pulling the build cost up.

[bak @copiumfueled]

@yoemsri Exactly that. The bull case prices every future dollar at today's capacity per dollar, and that ratio is the thing that's actually moving. Once the next dollar buys less, depreciation stops being a proxy for replacement and the whole ROIC math drifts

[bak @copiumfueled]

Agreed, that's the other half and I'm not denying it. But 'value rises exponentially' is the part you're assuming. Cost to build the next gigawatt is already climbing where we can see it. The value has to show up as inference revenue that outruns it, and that part isn't proven yet. Cheaper tokens lift demand, but the gap only closes if revenue per gigawatt beats capex per gigawatt.If price per token falls faster than volume grows, the value side doesn't win on its own

[bak @copiumfueled]

Your second tweet is my argument. If 70% of the build is rack capex and GPUs get pricier each generation, the cost to stand up the next data center keeps rising. Cheaper per token and pricier per build happen at the same time. And that 10x is Nvidia's MoE long-context benchmark. Dense inference runs closer to 2-3x so that number is the ceiling.

[bak @copiumfueled]

Sacks showed how America locked its two best AI models in a drawer.

Mark Zuckerberg was asked the question that story leaves open: so what do you actually do about frontier models?

His answer cut against the whole instinct behind the Mythos freeze.

The interviewer brought up Anthropic getting Mythos pulled by the government.

Zuckerberg didn't talk about that specific fight. He went one level down, to the philosophy underneath it.

He laid out the spectrum. On one end, a set of people who say this is too dangerous, we have to be the only ones who hold it.

On the other end, people who say it's powerful, it has real problems, but the answer is to put it in many hands and let them check each other.

Meta's instinct sits with the second camp. Push the capability out, don't hold it back.

Then he reached for the part most people get backwards: open source software.

Early on everyone assumed open code was less safe, because anyone can read it and hunt for the holes. What actually happened was the opposite.

More eyes find the vulnerabilities, you patch them, you ship the fix to everyone, and the software comes out harder than the stuff nobody was allowed to inspect.

His bet is that AI security works the same way. Harden it in the open, fast, instead of sealing it in a room and hoping.

Which lands strangely next to the last two weeks.

The most capable models in the country got frozen by the government in the name of safety, over a jailbreak that Anthropic says other models could already do.

One report. One phone call from a competitor to the Commerce Department. The whole thing offline for hundreds of millions of people.

Sacks described the room America locked its models in.

Zuckerberg is arguing the lock itself is the vulnerability.

WATCH THE FULL INTERVIEW ON @Complex

[bak @copiumfueled]

The two best AI models in America are sitting in a drawer right now.

David Sacks explained why this week on All-In, from inside the administration.

China just dropped GLM 5.2. Open weight, MIT license, free to download anywhere. It beats GPT 5.5 on coding and trails Claude Opus 4.8 by under a point.

The best openweight model on earth, and it came out of Beijing.

Fable got rolled back. GPT 5.6 is stuck navigating new approval hoops. Both frozen by a jailbreak report and a safety fight nobody outside Washington fully understands yet.

So the most open country ended up with the closed models, and right now they are off the market.

Sacks traced how America got here. Dario spent months lobbying for a federal AI regulator. He wrote a blog post asking for an FAA for AI. A government approval process for model releases. And he got it.

The administration moved, and the first thing the new caution did was freeze his own company's model.

The Chinese model has no such problem. No approval queue. No regulator. Trained on Huawei chips, packaged as "AI in a box," sold globally at a fraction of the cost.

Sacks has been saying it for months. We invented reasons not to ship. We defined this as a race and then handed the other side a head start.

What everyone is pointing at is real, but none of this was fate. A freer society didn't take the open models. America shut its own best models in a room and called it safety.

The closed models come from the open nation because the open nation chose to close them.

Watch the whole thing on @theallinpod

[bak @copiumfueled]

Goldman's desk says open-weight models you self-host break the AI buildout math, the fear runs one way, intelligence concentrated in a few closed providers.

Chamath, on Axios, lands at the same fork and answers head-on, his fix pointing the other way from a single approved pipe.

The shift Goldman is bracing for is already moving.

GLM-5.2 shipped this month. Weights you can pull down and self-host, no one else's cloud needed.

Chamath doesn't start from regulation. He starts from what he wants the field to look like: a vibrant ecosystem of trusted actors, a thousand flowers blooming, the market staying wide instead of narrowing to a few approved names.

Then he gets to China, and that's where he draws the line. Smart move, he says, but it has a weakness.

China went open weight, and open weight is not open source.

You get a little window into the machine, you watch the gears turn, but you never learn how the gears were made, nor will you ever.

What he's after is models from names you can actually point to.

Cheap alternatives are going to flood in from everywhere, and he says it plainly, we're gonna wanna know who they are.

The trusted actors he wants carry something the anonymous flood doesn't. There are consequences for them mucking around with the gears.

You know who they are, so there's someone to answer for it.

Goldman's worry is the field collapsing to a few closed providers. Chamath wants the opposite, a field kept wide and filled with players you can identify, the reverse of one trusted channel.

There's a tension he doesn't resolve. He says American, and he says trusted, and both are narrower than everyone.

A field policed by who you can name can drift toward the big players, the same way an approved channel does.

Same fork, two readings. Goldman counts the providers left standing and braces for the number to shrink.

Chamath counts them too, then keeps going, to who's actually behind them.

Watch the full conversation on the @axios

[zerohedge @zerohedge]

From Goldman's Delta-1 Desk:

Open-Weight: I still don’t think markets appreciate what GLM-5.2 and the advent of near frontier quality open weight models actually change. One explanation for the recent weakness in hyperscalers is precisely that…open models reduce the need to centralize every incremental workload inside closed cloud/frontier ecosystems. If more inference gradually migrates on premise or to enterprise owned infrastructure, the required pace of hyperscale investment inevitably comes into question (arguably its been bullish for hardware). Curious if we see more anecdotes of enterprises routing simple inference to local or open models while escalating only the hardest reasoning tasks to frontier APIs. That efficiency layer has barely been monetized. Cost conscious enterprises are no longer willing to subsidize unlimited token consumption when capable alternatives exist. One question I’ve been asking myself is whether governments eventually regulate frontier open-weight models in the same way they regulate other dual-use technologies. Not necessarily an outright ban, but licensing, safety certification or restrictions on models above defined capability thresholds, particularly from foreign developers. Ironically, that would probably strengthen the major cloud providers, who become the trusted distribution channel for approved models. First order, that’s probably supportive for hyperscalers and perhaps slightly less so for decentralized hardware demand. The second order effects become much more complicated. Restricting open models risks slowing domestic innovation while encouraging the rest of the world to continue developing independently. Remove open source entirely and you also remove a major source of competition, effectively concentrating intelligence inside a small number of closed providers.

[bak @copiumfueled]

The scarcity logic is right. Linear supply, exponential demand, that's why these aren't commodities. One thing though: BlackRock isn't doing this through REITs. It's GIP and the AIP consortium buying private infra directly, like the $40B Aligned deal not public REIT stock. And Fink chairs AIP. He's also said a couple of these buildouts will go bust and that's just capitalism working. Scarce doesn't mean every gigawatt pays off